Paul Byszkin

• • Made stateful FIX (QuickFIXJ) Trading gateways HA using Jgroups Raft (CP from CAP) on 2 ms latency to remove 1 hour recovery time and DevOps manual intervention.

• • Profiled (JFR and perf) for CPU and GC pressure and modernized Market Data gateways running on Java 17 with ZGC to use EMA Refinitiv instead of legacy InvisionJ to increase performance by 50% to 37k FIX refresh messages per second (5 level pricing ladder with daily session data).

• • Developed Pricing Refinitiv SDK-based load test simulator to automate load testing for protection against performance degradation over time and allow to test optimization ideas.

• • Worked with Pricing team to extend our Market Data offering from touch to 5 level ladder to protect us from excessive hedging costs (multi million savings).

• • Performance tuning of data structures and Avro schemas to reduce CPU and GC pressure in hot paths to avoid long stop the World pauses and stay under 1ms max.

• • Put multiple Git repositories under monorepo for easier development and faster CI-CD GitLab pipelines.

• • Scaled down Market Data stack to 50% of CPU and 25% of memory to reduce cost of infrastructure.

• • Added docker containers which allowed to decommissioned TEST environment due to no downstream or client dependencies.

• • Added very strict integration tests to cover upstream and downstream protocols with parallel execution in 1 minute to allow safer refactors and adding new features.

• • Implemented simple DevOps automation for our on-prem using Rust and expect while waiting for long term strategic solution provided by Platform Team.

• • PoC using QuickFIX with Rust to utilize SerDe, no nulls / exceptions / no GC on hot paths.

project.

• • Split and untangled huge Java 17 monolith codebase (SpringBoot) into manageable packages (ElasticSearch, Hadoop, HBase) to allow build modularization.

• • Optimized build (from 4h to 1h) by parallelizing and using in memory ElasticSearch

• • provisioned with minimum data set rather than connecting to TEST environment. This allowed for a much faster feedback loop.

• • Dockerized ElasticSearch to speed local development and simplify builds.

• • Quickly but temporarily added HA to the existing Tibco RV communication layer (partitioned network) to speed up Disaster Recovery.

• • Been asked to migrate off Tibco so run a set of PoCs and benchmarks (network latency, serialization cost, CPU usage, GC pressure) to find a long-term replacement.

• • Re-architectured whole middleware from broadcast to point-to-point with service discovery.

• • Prepared the whole stack to migration to cloud – k8s, docker, virtualization as part of company wide modernization.

• • Added a replication mechanism for caches based on existing stack and existing constraints.

• • Added symmetric and asymmetric (based on redirect - 302) load balancing.

• • Integrated HA clients with service discovery based on Eureka.

• • Critical changes were behind feature flags (some realtime) to reduce the risk.

• • Split 2GB git mono-repo (maintaining history) due to organizational change and size limit in BitBucket.